Scope & disclosure. As-of date 2026-10-18 (drafted 2026-08-04; every load-bearing claim re-verified against primary sources on 2026-10-18; see the currency note at the end). Editorial thesis, not investment advice. Disclosure: the sponsor is long NVIDIA (NVDA) and BlackBerry (BB), and holds no position in Tesla, Mercedes-Benz or Volvo. NVIDIA is this piece’s subject, and BlackBerry’s QNX is named below as one beneficiary layer of this thesis. I’ve named QNX’s two closest certified competitors right alongside it, and NVIDIA’s own bid for that layer too, so you can weigh the argument, not the position.
Executive Summary
The crowd read the headline as a funeral: NVIDIA just open-sourced its best self-driving model, so autonomy is now free and NVIDIA’s edge is gone. Here’s the overlooked angle. Giving away the model doesn’t give away the moat. It moves the margin down the stack. NVIDIA mailed every automaker a free jet-engine blueprint. Great. You still can’t fly it anywhere, because the FAA doesn’t certify blueprints. It certifies the plane.
Hold onto that image, because we’re going to build the whole argument on it. A blueprint is knowledge: the hard-won design of a thing that works. A certified plane is that knowledge turned into a machine that a regulator has signed off to carry human beings. The distance between those two is enormous, it is expensive, and it is exactly where the money lives. All NVIDIA gave away is the blueprint.
On 2026-08-04, NVIDIA confirmed that Alpamayo 2 Super, a 34-billion-parameter reasoning model for autonomous vehicles it first announced back on May 31 at GTC Taipei, is now cleared for commercial use under a permissive open license [1][2]. That’s the real news peg, not a rehash of the spring announcement. But NVIDIA itself calls this model a “teacher,” not a shippable product [3]. Between the open weights and a car that legally drives sit two toll booths: the distillation tooling that shrinks the model to real-time size (mostly NVIDIA’s own), and the safety-certified runtime the shrunk model has to run on (mostly not NVIDIA’s).
So the falsifiable claim: if you’re pricing “who wins as self-driving software gets commoditized,” look at the certified-runtime layer (QNX, Green Hills, Wind River), not the model layer. Free brain. Licensed highway. The rest of this piece explains, in plain English, why the brain being free is exactly what makes the highway valuable.
Key Findings
- Alpamayo 2 Super is a 34B-parameter vision-language-action (VLA) model, a 32B reasoner (“NVIDIA Cosmos 3 Super Reasoner”) bolted to a 2B action expert, announced May 31 and cleared for commercial use on 2026-08-04 under the OpenMDW-1.1 open license [1][2][4]. A VLA is a model that takes in what the car sees, reasons about it in language, and outputs an action (a steering trajectory).
- NVIDIA designed it as a “teacher model,” not an in-car product. In its own words, it “can be distilled into compact models that run on NVIDIA DRIVE AGX Thor inside the vehicle” [3]. Open weights are a curriculum, not a driver.
- The mechanical reason it can’t drive as-is: latency. Independent research (Carnegie Mellon) clocks raw reasoning-model driving decisions at 500ms–2s or worse, against a control-loop budget of roughly 50–100ms, the pace a car actually needs to make decisions. Distillation closes that gap, with measured speedups of 7.9× to 44.8× [5].
- No open model clears the safety bar, because that bar isn’t a model test. ISO 26262 (ASIL-D), ISO 21448 (SOTIF), the new ISO/PAS 8800 for AI, and UNECE R155/R156 certify an integrated system and its lifecycle, not a set of weights [6][7]. The small set of vendors with runtimes already through that gate includes QNX (275M+ vehicles, a royalty backlog of about $950 million at fiscal 2026 year-end, plus an “over $100 million” design win announced since) alongside Green Hills INTEGRITY and Wind River VxWorks [8][9][10][11].
- The value-capture picture is not one-sided. Tesla’s vertical integration and China’s Horizon Robotics (the country’s #2 assisted-driving chipmaker, ~13.6% of the total China market behind NVIDIA) both route around parts of this stack [12][13].
🧠 What’s actually going on
Let me define the thing first, because the plain-English version is where the crowd’s error hides, and because you can’t reason about who gets paid until you know what the product actually is.
Start at the bottom. A modern AI model is not a program someone wrote line by line. It’s a giant tangle of numbers, call them parameters, that were tuned, not authored, by showing the system millions of examples until it learned the statistical shape of the task. That tangle is a neural network: layers of simple math units, loosely inspired by neurons, where each connection carries one of those tunable numbers. When people say Alpamayo 2 Super is “34 billion parameters,” they mean the network has 34 billion of those adjustable dials, all set during training [1]. The dials are the knowledge. Which is why this next distinction matters so much.
“Open weights” is not the same as “open source,” and the gap is the whole story. In traditional software, “open source” means you get the human-readable recipe: the code, which you can read, change, and recompile. With an AI model, the “recipe” that produced it (the training data, the training runs, the exact process) usually stays private. What gets released are the weights: the finished 34-billion-number tangle itself. You get the trained brain, but not the notebook that explains how it was raised. That’s still hugely useful (you can run it, study its behavior, and build on it), but it’s a result handed to you, not a method you can fully reconstruct. Open weights are the artifact. Keep that straight and the rest of the argument clicks into place.
Now, what does this particular artifact do? Alpamayo 2 Super is a vision-language-action (VLA) model. Break that jargon down into its three verbs, because that’s literally how it runs:
- Vision. It takes in what the car’s cameras see. Raw pixels in.
- Language. It reasons about that scene in something like words. NVIDIA calls its version a “Chain-of-Causation” trace: the model’s stated why: “there’s a construction cone, the lane is closing, yield to the merging truck.” This is the same trick that made chatbots suddenly seem smart: forcing the model to “think out loud” in steps rather than blurting an answer.
- Action. It outputs a driving trajectory: the actual path and speed the car should take. Language reasoning in, steering out.
Architecturally that’s a 32B reasoner paired with a 2B diffusion-based action expert, post-trained with reinforcement learning [1]. You don’t need those last terms to follow the argument; the point is that a big part (32B) does the thinking and a small part (2B) converts the thinking into a maneuver. NVIDIA aimed the whole thing squarely at the long tail: not everyday cruising, which is largely solved, but “the rare, complex situations”: construction zones, partially occluded pedestrians, weird four-way-stop standoffs [2]. In a driving system, the ordinary miles are easy and the one-in-a-million weird moment is what kills people, so the long tail is exactly the part worth the reasoning.
The date discipline matters, because sloppy coverage is blurring two events. Alpamayo 2 Super was announced on May 31 at GTC Taipei. What happened today, 2026-08-04, is different and specific: NVIDIA confirmed the model is “now available for commercial use” on Hugging Face under OpenMDW-1.1, the Linux Foundation’s new permissive license built for open AI model distributions [2][4]. “Permissive” here means the license lets you fine-tune it, build derivative models, and even redistribute commercially without paying NVIDIA a royalty on the weights. An automaker can legally build on it. That is a genuinely open release, and it is genuinely a big deal for the R&D floor: it means a company starting from zero no longer has to spend years and a fortune training a frontier driving brain from scratch.
But here’s NVIDIA’s own framing, and it is the whole story in one sentence: “As a teacher model, it can be distilled into compact models that run on NVIDIA DRIVE AGX Thor inside the vehicle.” [3] Read that again. NVIDIA is not handing you a driver. It’s handing you a teacher: a big, slow, cloud-scale model whose job is to train a smaller one. Think of a world-class professor: brilliant, but you can’t fit them inside a dashboard, and they can’t answer in the tenth of a second a moving car demands. The professor’s job is to produce a fast, cheap graduate who can sit in the car. That graduate, the “student” model, is what ships. This is the first layer of NVIDIA’s own physical AI stack, and the model is the cheapest thing in it, precisely because it’s now free.
Actionable Takeaway: open weights ≠ a shippable product. When you see “Company X open-sourced its AI,” the right question is never “so it’s free now?” It’s “what’s the next gate, and who owns it?”
🚀 The overlooked play — the two toll booths
Between a 34B open teacher and a car that legally carries passengers, there are two places somebody collects a toll. Neither is the model. Let’s walk through both, and why each one is a toll booth rather than an open road.
Toll booth one — distillation (mostly NVIDIA’s). Distillation is the process of taking a big, slow “teacher” model and training a small, fast “student” model to imitate it: you keep most of the smarts, you shed most of the compute. Here’s the mechanism in plain terms: you run the giant teacher over huge amounts of driving data and record not just its final answers but the reasoning and trajectories it produces. Then you train a much smaller network to reproduce those same outputs directly. The student never has to think as hard, because it’s not deriving the answer from first principles; it’s learning to mimic a brain that already did. The result is a model with a fraction of the parameters that behaves, most of the time, almost like the big one.
Why is this not optional? Because of a hard fact about how these networks run. Every time the model makes a decision, the machine has to push the input through all of those parameters, doing billions of multiplications. More parameters means more math means more milliseconds. A 34B teacher is smart precisely because it’s big, and slow for exactly the same reason. And a car cannot wait.
That’s where the second concept comes in: the real-time control loop. A self-driving car isn’t a chatbot that can pause to think. It runs a loop (sense the world, decide, act) over and over, many times per second, like a drummer who cannot miss a beat. Independent researchers at Carnegie Mellon measured raw reasoning-model driving decisions at 500ms to 2 seconds (some setups near 5s), while safe closed-loop vehicle control needs a fresh decision roughly every 50–100ms, call it 10 to 20 times a second [5]. Put those numbers side by side: at highway speed, a car covers real ground in the half-second-to-two-seconds the raw teacher spends thinking. That’s not a rounding error; that’s the difference between stopping in time and not. As the CMU paper puts it, “slow inferencing leads to unsafe delays in trajectory updates, reduced responsiveness and inefficient decision making in time-sensitive contexts.” [5] Their fix, distillation, bought a 7.9× to 44.8× speedup turning a large VLA teacher into a real-time-viable student [5].
So the gap between the open teacher and a shippable student is a factor of ten-to-forty, and closing it is skilled, tooling-heavy work, and the tooling that does it is NVIDIA’s own: AlpaSim and AlpaGym (for running the student through millions of simulated miles), the Halos safety-validation workflow, Omniverse NuRec (which turns real fleet footage into synthetic training scenes), the Physical AI Open Datasets, and the DRIVE AGX Thor / Hyperion silicon it all targets [1][3]. NVIDIA gave away the curriculum and kept the school. The blueprint is free; the machine shop that turns it into a running engine is still on the meter. That’s toll booth one, and notice it’s a toll on effort and tooling, not on a license fee.
(Exhibit 2, the teacher→student→certified-runtime schematic, appears at the end of this section.)
Toll booth two — the certified runtime (not NVIDIA). Now say you’ve distilled a beautiful little student model that hits the latency budget. You still can’t sell a car with it. Here’s the part almost nobody outside the industry understands: a driving model isn’t a legal product until the system it runs inside clears a stack of safety bars, and those bars don’t test the weights. They certify an integrated system, its documented safety case, and the entire process used to build it. Let me make that concrete, because “certification” sounds like a rubber stamp and it is the opposite of one.
Functional-safety certification is the discipline of proving, on paper and in test, that a system won’t hurt someone when something goes wrong. In cars it’s governed by ISO 26262, and its top integrity rung is ASIL-D, the Automotive Safety Integrity Level reserved for the functions that kill people if they fail (think steering and braking). Earning ASIL-D is not a weekend audit. It means building the software under a controlled process from day one, documenting every requirement and how it’s tested, proving your development tools themselves won’t inject faults, and having an independent assessor (a body like TÜV) verify the whole chain. That’s years of work and a paper trail thousands of pages deep, which is why so few vendors have it, and why you can’t retrofit it onto a model you downloaded this morning.
And ISO 26262 is only the start. ISO 21448 (SOTIF), Safety Of The Intended Functionality, covers a different, AI-shaped problem: the system works exactly as built and still makes the wrong call, because the world showed it something its training didn’t cover. That’s the perception-and-planning gap ISO 26262 never reaches, and it’s precisely where a driving AI is most dangerous. ISO/PAS 8800, published only recently, extends functional-safety practice to AI components specifically [6]. On top of those sit the binding UN vehicle-approval rules UNECE R155 and R156, which require a certified process for cybersecurity and for software updates across the vehicle’s entire lifecycle, because a car that gets over-the-air updates for ten years has to prove every future patch is safe too [7]. Open model weights satisfy exactly none of this. They can’t, by their nature: a bag of numbers has no process, no safety case, and no lifecycle.
This is the deepest point in the whole piece, so let me put it in the metaphor. The certified runtime is the airframe and the FAA type-certificate together: the physical machine the engine bolts into, plus the years-long regulatory proof that the combination is safe to carry passengers. NVIDIA can mail you the world’s best engine blueprint for free. It does nothing to shorten the certification queue, because certification isn’t a test of the engine’s cleverness. It’s a test of the whole aircraft’s discipline.
The vendors who have runtimes through that gate are few, and this is where the margin pools. Before we name them, one more plain-English term: a royalty backlog. When a car company designs a vendor’s certified operating system into a future model, it commits to pay a small per-vehicle royalty for every unit that eventually ships, often years out. The “backlog” is the sum of those future, contracted payments. It’s a forward-looking measure of demand already locked in, which is why it’s such a good tell for where value is accruing.
QNX is the loudest name. Its safety OS, hypervisor, and virtualization stack are certified or pre-certified to IEC 61508 SIL 3, ISO 26262 ASIL-D, IEC 62304 Class C and ISO/SAE 21434 [8]; its hypervisor was the world’s first ASIL-D-certified commercial hypervisor, back in 2019 [9] (a hypervisor, if it’s new to you, is the software that lets one chip safely run several isolated systems at once, say the safety-critical driving stack walled off from the infotainment screen); and it sits in 275M+ vehicles, with a royalty backlog of about $950M at the end of fiscal 2026 (February 2026) and design wins at all top-10 automakers [10]. In September it added a single win, from Coretura (the Volvo Group and Daimler Truck joint venture), that BlackBerry says adds “over $100 million” to that backlog [16]. But QNX is not alone, and this thesis doesn’t need it to be. Green Hills Software’s INTEGRITY is ASIL-D certified and picked up an AISIN driver-monitoring win in 2026; Wind River’s VxWorks Cert Edition is separately ASIL-D / SIL 3 certified [11]. Three credible toll collectors, not one. That’s the point: the certified-runtime layer is a small, gated club, membership takes years to earn, and nobody is open-sourcing their seat.
Who does this actually affect? Everybody downstream of the model. A robotaxi startup that cheered the free weights still has to license a certified runtime or spend years building its own. A legacy automaker that wants to adopt Alpamayo inherits the exact same gate. A Tier-1 supplier assembling the compute module has to pick a certified OS to run it on. The rider, at the end of the chain, is only allowed in the car because that gate exists. Free weights lower the cost of the brain to roughly zero, which just makes the scarce thing, the certified place to run it, relatively more valuable, not less.
Exhibits
Two figures anchor this section. Read them together. Exhibit 2, Teacher → Student: the open weights are the teacher; what ships is the student, and the certification bar applies to the student’s deployment, not the teacher’s weights. Exhibit 1, Where the value sits: the 6-layer autonomy stack, commoditized at the bottom, gated in the middle (the certified-runtime chokepoint), OEM-owned at the top. The open model is the layer with the least defensibility.
Actionable Takeaway: if you’re trying to price who benefits as self-driving software commoditizes, track the certified-runtime layer’s design wins and backlog (QNX’s, Green Hills’, Wind River’s), not the model-license news. The press release is the distraction. The backlog is the plumbing.
⚠️ Risks & the counter-case
I’d be doing you a disservice if I sold this as a sure thing. Here are the four strongest arguments against me, steelmanned then answered. If you only skim one section, skim this one, because a thesis is only as good as the case against it.
1. The commoditization case isn’t fully wrong. Open-sourcing genuinely lowers the R&D floor. That’s the whole reason it’s news. An automaker or Tier-1 with no prior autonomy path now has one, and that’s a real threat to NVIDIA’s model-licensing revenue specifically. Here’s the twist: the stronger version of this objection actually helps my conclusion. Imagine NVIDIA’s tooling gets commoditized too: a credible open rival to Omniverse NuRec appears, or someone ports the distillation pipeline onto a cheaper accelerator. Then layers 1, 2, and 3 of the stack (weights, tooling, silicon) all collapse toward free, and the certified-OS layer becomes the only remaining moat. That would break NVIDIA’s slice of the story while validating the runtime conclusion even harder. The moat gets narrower and deeper, not gone. Notice what that means for an investor: “NVIDIA loses” and “the certified-runtime thesis wins” are not opposites. They can be true at the same time.
2. Vertical integration bypasses the whole framework. See Tesla. Everything above assumes an automaker buys the pieces from different vendors: NVIDIA’s model, NVIDIA’s tooling, someone’s certified OS. A company big enough to build all of it in-house pays none of those tolls. Tesla runs its own custom AI silicon and its own end-to-end stack, and depends on NVIDIA DRIVE hardware for exactly none of its shipping cars [12]. That’s a live counter-example, and I won’t wave it away: this thesis applies to the NVIDIA-dependent middle of the market, meaning most legacy OEMs, Tier-1 suppliers, and robotaxi startups without a Tesla-scale balance sheet, not to everyone. State the scope plainly and the argument holds; overstate it and Tesla breaks it. The lesson generalizes: any toll booth can be bypassed by whoever can afford to build their own road, so the moat is real only for those who can’t.
3. The silicon layer is regionally contestable, and it is the weakest leg. A moat also fails if a government simply legislates around it. In China, domestic chipmaker Horizon Robotics has risen to the country’s #2 assisted-driving chipmaker behind NVIDIA, at roughly 13.6% of the total China market as of April 2026 (its ~49% figure is share of domestic suppliers only, not the whole market), and Volkswagen has partnered with Horizon for its China driver-assist systems [13]. Policy-driven localization, not a technology loss, is doing the work. Beijing wants a home-grown stack, so the market bends. Note precisely what this dents: it weakens the silicon leg of the stack, not the certified-OS leg. China runs its own domestic safety-certification apparatus, and QNX and its competitors still have to win those cycles separately. But “NVIDIA silicon = permanent global toll booth” is the claim I’d hedge hardest. The certified-runtime conclusion survives this bear case; the silicon-supremacy assumption does not.
4. “Reasoning” is vendor framing until proven otherwise. This is the honest caveat, and it’s a safety point, not a nitpick. NVIDIA’s long-tail improvement claims rest on its own benchmark: 1,434 challenging samples and 913 reconstructed scenes, on NVIDIA’s own eval set, not independently replicated [2]. And there’s a first-principles catch worth naming, because it’s a genuine open problem in AI: a reasoning trace isn’t guaranteed to be faithful. A model can emit a plausible-looking “why” that reads like careful thought but doesn’t actually reflect how it reached the decision. The explanation is generated alongside the answer, not necessarily from the same process. Which means a convincing explanation is not the same as a correct one, and for a safety system that gap is the whole ballgame. Treat “it reasons about construction zones” as a marketed capability pending third-party validation, not a settled fact.
🍌🐀 Read: Want the pipeline made concrete? Back at CES in January, the Mercedes-Benz CLA was held up as the first production car shipping NVIDIA’s full AV stack with Alpamayo reasoning, targeted for Q1 2026 [14]. That’s the teacher→student→certified-runtime→shipped-car chain happening in metal, not a whiteboard: a cloud-trained brain, distilled down to fit a dashboard, running on a certified system, inside a car a regulator let onto the road. One honest asterisk: that reporting predates the “2 Super” release and refers to the earlier Alpamayo generation, so read the CLA as color for the pattern, not proof that Alpamayo 2 Super specifically is in that vehicle.
The falsifying signal. In a base case (moderate confidence), I’d expect certified-runtime backlogs to accelerate among new Alpamayo-adopting OEMs over the next 2–3 quarters as these stacks move from announcement to shipping metal. More free brains in circulation should mean more demand for certified places to run them. The bull case (lower confidence) is that NVIDIA’s tooling becomes the reference pipeline the way CUDA did, and the certified-OS layer becomes a value-share partner rather than a rival. The bear case is Tesla and China above. Here’s what would prove me wrong: if Alpamayo-adopting OEMs increasingly self-certify or route through lighter compliance paths instead of licensing a certified runtime, the toll booth leaks and the thesis fails. That’s the honest test. I’ve told you the number that would sink my own argument.
Verdict — who actually captures the value
So who gets paid? Walk the stack one layer at a time, and ask the only question that matters at each floor: is this scarce, or is this free?
Layers 1 through 3 (the open weights, the distillation tooling, the silicon) trend NVIDIA-adjacent but contestable. The model is now free by NVIDIA’s own choice. The tooling is proprietary today but could be cloned. The silicon is real leverage but is already being routed around in China. None of those is a permanent lock. Layer 4, the safety-certified runtime, is different in kind: it’s a durable gated chokepoint for the NVIDIA-dependent middle of the market: a club of maybe three qualified vendors, each of whom spent years and fortunes earning a certification you cannot download, and none of whom is open-sourcing their seat. Layers 5 and 6 (final integration, the safety case, and legal liability) stay with the automaker, where they always were, because in the end it’s the OEM’s name on the type-approval and the OEM’s neck on the line.
One honest wrinkle, because it is the version of my wrong-case I’d watch hardest: NVIDIA does not leave that fourth layer empty. It sells its own safety software, Halos OS, which it says is built on DriveOS, certified ASIL-D conformant, with a hypervisor that walls safety-critical functions off from the AI workloads [18]. And when Volvo Cars widened its NVIDIA partnership on 2026-10-15, the package it named was Hyperion, Halos OS and the Alpamayo Stack: the brain, the box and the safety operating system, all from one vendor [17]. If automakers start buying the certified runtime from NVIDIA instead of from QNX, Green Hills or Wind River, the toll booth doesn’t disappear. It moves, and it moves toward NVIDIA. That would still prove the thesis that the value sits in the certified layer, but it would change who collects.
That’s the metaphor paying off in full. NVIDIA mailed everyone a free jet-engine blueprint, and the internet cheered that flight just got free. But the FAA doesn’t certify blueprints. It certifies the plane: the airframe the engine bolts into and the mountain of safety evidence behind it. The blueprint is the model. The certified plane is the runtime plus the safety case. Making the blueprint free doesn’t make the plane free; if anything, it draws a brighter line around the part that still costs years to build.
This is not hype. It’s the boring mechanics of how a demo becomes a legal product (sense, decide, act, ten times a second, on a system a regulator has signed for), and the boring layer is the one with the toll booth.
The tell — what to watch
Since this piece was drafted, the distribution half of the thesis has already shown up in public numbers. NVIDIA reported on 2026-08-04 that Alpamayo had passed 500,000 downloads on Hugging Face, which it calls “the most-adopted open reasoning model family for autonomous driving” on the platform, and it has extended the OpenMDW licence across the entire Alpamayo family, the earlier 10B-parameter Alpamayo 1 and 1.5 releases included, so any of them can be deployed commercially without further permission [15]. That is the free brain spreading faster and wider than the single 34B release this post started from. It settles the first half of the argument and leaves the second half exactly where I put it: if the model is now everywhere and free, the question of who gets paid moves entirely to the layer that can legally carry a passenger.
One number, publicly reported, tells you whether this thesis is working: the royalty-backlog and design-win trajectory at the certified-runtime vendors, of which QNX’s is the most visible [10], over the next 2–3 quarters. The first datapoints lean the way this thesis predicts, but they are not the test. BlackBerry reported on 2026-09-24 that QNX revenue hit a record $80.3M, up 27% from $63.1M a year earlier, and that one new design win, Coretura’s, adds “over $100 million” to the royalty backlog [16]. BlackBerry did not publish an updated backlog total, so the ~$950M year-end figure is still the last hard number, and Coretura is a commercial-truck program, not an Alpamayo adopter. That is demand for certified runtimes in general, not yet proof that Alpamayo-adopting OEMs are paying the toll. The next read is BlackBerry’s next quarterly report. Remember what that backlog is: contracted future per-vehicle payments from cars that haven’t shipped yet, i.e. demand already committed. If it accelerates specifically among new Alpamayo-adopting OEMs, that’s the toll booth collecting: free brains creating paying traffic on the licensed highway. If OEMs building on Alpamayo instead self-certify or take lighter compliance paths, it isn’t, and I’ll have been wrong in the exact way I told you to check. Watch the plumbing, not the press release.
Giving away the model didn’t give away the moat. It moved the toll booth from the thing everyone was staring at to the layer nobody was pricing. If you want the full map of that layer, the quiet certified runtime the whole physical-AI age has to run on, read the physical-AI thesis →.
The 🍌🐀 has spoken. 🍌🐀
Subscribe for the next overlooked play. Banana Rat research.
Sources
[1] “NVIDIA Launches Alpamayo 2 Super Open Reasoning Model for Robotaxis,” NVIDIA Newsroom, 2026-05-31 (announced at GTC Taipei). https://nvidianews.nvidia.com/news/nvidia-alpamayo-2-super-robotaxis
[2] “NVIDIA Alpamayo 2 Super, the Frontier Open Model for Robotaxis and Autonomous Vehicles, Now Available for Commercial Use,” NVIDIA Blog (Jessica Soares), 2026-08-04. https://blogs.nvidia.com/blog/alpamayo-2-super-open-model-now-available/
[3] “Generate Trajectories, Reasoning Traces and Auto-Labels with NVIDIA Alpamayo 2 Super,” NVIDIA Developer Blog, 2026-08-04. https://developer.nvidia.com/blog/generate-trajectories-reasoning-traces-and-auto-labels-with-nvidia-alpamayo-2-super/
[4] “Linux Foundation Releases OpenMDW-1.1; NVIDIA Adopts OpenMDW for Cosmos, Isaac GR00T, Ising and Nemotron AI Model Families,” Linux Foundation press release, 2026-05-28. https://www.linuxfoundation.org/press/linux-foundation-releases-openmdw-1.1-nvidia-adopts-openmdw-for-cosmos-isaac-gr00t-ising-and-nemotron-ai-model-families
[5] Huang, Hua, Zhou, Sural, Rajkumar, “RT-VLA: Real-Time Vision-Language-Action Models via Knowledge Distillation,” Carnegie Mellon University, arXiv:2606.14010, submitted 2026-06-12. https://arxiv.org/html/2606.14010
[6] “Introducing ISO/PAS 8800 – Functional Safety for AI in Road Vehicles,” SGS, ~2025-04 (ISO 21448 / SOTIF published 2022; standards facts). https://www.sgs.com/en/news/2025/04/safeguards-04625-introducing-iso-pas-8800-functional-safety-for-ai-in-road-vehicles
[7] “Software Update Management Systems According to UNECE R156,” UL Solutions (UNECE R155/R156 in force since 2021-01). https://www.ul.com/sis/insights/software-update-management-systems-according-unece-r156
[8] “Certifications & Compliance,” QNX official page, current as of 2026-08-04. https://qnx.software/en/developers/developer-roadmap/certifications-and-compliance
[9] “BlackBerry QNX Hypervisor Awarded World’s First Automotive Safety Integrity Level (ASIL) ‘D’ Certification,” PR Newswire, 2019. https://www.prnewswire.com/news-releases/blackberry-qnx-hypervisor-awarded-worlds-first-automotive-safety-integrity-level-asil-d-certification-300969956.html
[10] “BlackBerry Reports Fourth Quarter and Full Fiscal Year 2026 Results,” 2026-04-09 (QNX in 275M+ vehicles, royalty backlog about $950M at FY2026 year-end; design wins at all top-10 automakers). https://www.nasdaq.com/press-release/blackberry-reports-fourth-quarter-and-full-fiscal-year-2026-results-2026-04-09
[11] Green Hills Software (INTEGRITY / Apex.OS ASIL-D) + Las Vegas Sun (AISIN selects Green Hills, 2026-05-26) + Wind River VxWorks Cert Edition (TÜV SÜD). https://ghs.com/news/20220105_CES_Integrity_Apex_OS_ASIL_D_driving.html · https://lasvegassun.com/news/2026/may/26/aisin-selects-green-hills-software-for-its-ai-powe/ · https://www.windriver.com/resource/tuv-vxworks
[12] “Clash of Self-Driving Technologies: Tesla vs. Nvidia,” ETC Journal / Yahoo Finance, 2026-01 (Tesla custom AI4/AI5 silicon; secondary/aggregator — general fact solid, chip-generation detail lower confidence). https://etcjournal.com/2026/01/07/clash-of-self-driving-technologies-tesla-vs-nvidia-january-2026/
[13] “Nvidia leads China assisted-driving chip market; Horizon Robotics rises to second,” Digitimes, 2026-06-18 (Horizon ~13.6% of total China market, #2 behind NVIDIA; ~49% is domestic-supplier share only). https://www.digitimes.com/news/a20260618VL211/nvidia-horizon-robotics-autonomous-driving-china-market-2026.html
[14] “NVIDIA Unveils Open-Source AI for Autonomous Driving, Ships in Mercedes-Benz CLA in Q1 2026,” Electrek, 2026-01-05 (illustrative color; refers to the earlier Alpamayo generation, not Alpamayo 2 Super). https://electrek.co/2026/01/05/nvidia-unveils-open-source-ai-for-autonomous-driving-ships-in-mercedes-benz-cla-in-q1-2026/
[15] “NVIDIA Alpamayo 2 Super, the Frontier Open Model for Robotaxis and Autonomous Vehicles, Now Available for Commercial Use,” NVIDIA blog, read 2026-09-10 (verbatim: “Alpamayo has already surpassed 500,000 downloads on Hugging Face, reinforcing its position as the most-adopted open reasoning model family for autonomous driving on the platform”; and “The OpenMDW license is now being applied across the entire Alpamayo model family”). Primary source — NVIDIA’s own.
[16] “BlackBerry Reports Second Quarter Fiscal Year 2027 Results,” 2026-09-24 (QNX revenue $80.3M, +27% year on year; Coretura Alloy Kore win “adds over $100 million to the QNX royalty backlog”; no backlog total given). https://www.newswire.com/news/blackberry-reports-second-quarter-fiscal-year-2027-results
[17] “Volvo Cars and NVIDIA Expand Partnership to Advance Safe Automated Driving,” NVIDIA newsroom, 2026-10-15 (“These include the Hyperion sensor and compute architecture, Halos OS, and Alpamayo Stack driving software”). https://nvidianews.nvidia.com/news/volvo-cars-and-nvidia-expand-partnership-to-advance-safe-automated-driving
[18] “Full-Stack Safety for Robotaxis and Autonomous Vehicles” (NVIDIA Halos), NVIDIA AI Trust Center, read 2026-10-18 (Halos Core “built on ISO 26262 ASIL D certified DriveOS”, with “a hypervisor that isolates safety-critical functions from AI workloads”; DriveOS 6.0 “is certified ASIL D conformant”). NVIDIA’s own claim. https://www.nvidia.com/en-us/ai-trust-center/halos/autonomous-vehicles/
Currency note (2026-09-10). Drafted 2026-08-04 and held. Before publication every load-bearing claim was re-checked against primary sources: the 34B architecture (32B Cosmos 3 Super Reasoner + 2B Action Expert) confirmed against NVIDIA’s own model card at huggingface.co/nvidia/Alpamayo2-Super and its developer-blog write-up, where the 32B + 2B split is stated explicitly — worth naming, because secondary coverage of this model is inconsistent (one scrape said 24B, and the widely-cited “10B” figure belongs to Alpamayo 1, a different and earlier model); the OpenMDW-1.1 licence confirmed, and since widened to the whole family; BlackBerry’s 275M+ vehicles confirmed; the royalty backlog was recorded as “$1 billion (Q1 FY2027)” in this pass; see the 2026-10-18 note below, which reverts it. Nothing came back wrong. One claim circulating in secondary coverage — that named OEMs including BYD, Geely, Isuzu and Nissan have adopted earlier Alpamayo models — is not stated in NVIDIA’s own materials and is deliberately omitted here rather than repeated. The download and licence figures in “The tell” were added in this pass.
Currency note (2026-10-18). Re-checked before publication. The “$1 billion” backlog recorded on 2026-09-10 could not be traced to a primary source (the cited release predates that quarter, and BlackBerry’s later releases give no total), so it is reverted to the last hard figure, about $950M at FY2026 year-end, with September’s Coretura win (“over $100 million”) noted separately [16]. “The tell” now reports BlackBerry’s 2026-09-24 quarter instead of looking forward to it, and the 500,000-downloads figure is dated to NVIDIA’s 2026-08-04 post. The 34B architecture, the OpenMDW-1.1 licence, the Horizon share figure and the Tesla line were re-confirmed. A concession paragraph was added after the layer verdict: NVIDIA sells its own certified safety OS (Halos OS on DriveOS), and Volvo Cars named it on 2026-10-15; both sourced to NVIDIA’s own pages [17][18].